Research On Wideband Adaptive Predistortion And Power Amplifier Behavior Model Based On Deep Learning

Power Amplifier(PA)is one of the core devices of the transmitter in the communication link,which amplifies the signal to improve the anti-interference ability of signal transmission and is also the main cause of signal non-linear distortion.Digital Predistortion(DPD)is widely used as one of the most promising linearization techniques,and the development of communication technology has put forward higher requirements for it.In this thesis,we focus on adaptive pre-distortion technology and deep learning based amplifier behavior model,mainly optimize the structure of pre-distortion device and reduce the computational complexity and improve the problem of lack of amplifier model accuracy under strong non-linear conditions,the specific work is as follows:Firstly,for the cost problem brought by the requirement of high speed ADC for the adaptive pre-distorter working under large bandwidth signal,this thesis uses a direct learning structure pre-distorter based on a single feedback signal for improvement.The experimental simulation results show that the method can reduce the number of ADCs on the feedback loop to one by optimizing the structure of the pre-distorter under the condition of ensuring the linearization effect of the pre-distorter,which can also achieve the purpose of reducing noise interference on the feedback link while reducing the cost.Secondly,for the problem of high computational complexity caused by the existence of large-scale matrix operations in the parameter iterative calculation process of the adaptive predistorter,this thesis combines the sample selection method based on QR decomposition with the direct learning structure predistorter.By extracting a set of representative small samples from the signals on the DPD feedback link,this method achieves the purpose of reducing the computational complexity without affecting the pre-distortion effect,and also avoids the problem of pathological system of equations for solving DPD parameters due to the small number of signal samples.In addition,this thesis also applies this QR decomposition-based sample selection method to the direct learning pre-distortion device based on a single real-valued feedback signal to further reduce the computational effort of DPD parameter solving while optimizing the DPD system architecture.Third,to address the problem that the strong nonlinearity of amplifiers in large bandwidth and high peak-to-average ratio scenarios leads to the lack of fitting accuracy of amplifier behavior models based on Volterra stages,a recurrent neural network DARNN amplifier behavior model based on a two-stage attention mechanism is proposed in this thesis.The model improves the behavioral model accuracy by the consideration of the attention mechanism,which considers the different memory depth input signals for the current output signal distortion with different effect sizes.In addition,this thesis analyzes the actual characteristics of the output signal nonlinear distortion which is mainly excited by the input signal envelope,and proposes a enhanced recurrent neural network A-DARNN amplifier behavior model with two-stage attention mechanism by cascading the nonlinear order terms of the input signal amplitude into the input signal vector,which further improves the amplifier behavior model accuracy.